Case study
Dataset for autopilot
Labeled dataset of road signs for training the autopilot that won the competition
Open project ↗01 Task
There was not enough data in the required class composition to train the traffic sign detector; the open sets did not match the markings and set of signs
02 Solution
I collected and labeled a dataset of road signs, selected images, brought them to a unified format, and annotated them by class. The structure was made to fit a standard learning pipeline so that the data could be picked up without changing the markup
03 Solved tasks
- Collected and selected images of road signs for the required composition of classes
- Brought images to a single format
- Added annotations by class
- Organized the structure for a standard training pipeline
- Trained a semi-automatic marker
- Using the marker, I doubled the dataset
- Completed the final processing and handed it over to the customer.
04 Result
A ready-made set on which the character detector is immediately trained. The composition of classes is expanded as needed. The autopilot trained on the basis of this dataset won the competition: